Guillermo Bustos-Pérez \(^{1,2,3}\), Javier Baena \(^{1}\), Manuel Vaquero \(^{2,3}\)
\(^1\) Universidad Autónoma de Madrid. Departamento de Prehistoria y Arqueología, Campus de Cantoblanco, 28049 Madrid, Spain
\(^2\) Institut Català de Paleoecologia Humana i Evolució Social (IPHES), Zona Educacional 4, Campus Sescelades URV (Edifici W3), 43007 Tarragona, Spain
\(^3\) Universitat Rovira i Virgili, Departament d’Història i Història de l’Art, Avinguda de Catalunya 35, 43002 Tarragona, Spain
Abstract
Lithic artefacts are usually associated with the different knapping methods used in their production. Flakes exhibit metric and technological features representative of the flaking method used to detach them. However, lithic production is a dynamic process in which discrete methods can be blurred, and in which features can vary throughout the process. An intermediate knapping method between the discoid and Levallois is commonly referred to under an umbrella of terms (the present research uses the term hierarchical discoid), and is associated with a broad geographical and chronological distribution throughout the Early and Middle Palaeolithic. This intermediate knapping strategy exhibits features of both the discoid and Levallois knapping methods, raising the question of the extent to which flakes from the three knapping methods can be differentiated and, when one is mistaken for another, the direction of confusion. An experimental assemblage of flakes detached by means of the three methods was used along with an attribute analysis and machine learning models in an effort to identify the knapping methods employed. In general, our results were able to very effectively differentiate between the three knapping methods when a support vector machine with polynomial kernel was used. Our results also underscored the singularity of flakes detached by means of Levallois reduction sequences, which yielded outstanding identification values, and were rarely erroneously attributed to either of the other two knapping methods studied. Mistaking the products of the discoid and hierarchical discoid methods was the most common direction of confusion, although a good identification value was achieved for discoid flakes and an acceptable value for hierarchical discoid flakes. This shows the potential applicability of machine learning models in combination with attribute analysis for the identification of these knapping methods among flakes.
Keywords: lithic technology; experimental archaeology; Levallois; discoid; Middle Palaeolithic; machine learning
Extended abstract
La producción de lascas se asocia a diferentes métodos de talla. Las lascas resultantes presentan características métricas y atributos que son representativos del método de talla del que se han producido. Sin embargo, la talla lítica es un proceso dinámico en el que los métodos de talla definidos pueden verse entremezclados debido a adaptaciones a las características volumétricas y de calidad de la materia prima, diferentes fases a lo largo del proceso de reducción, aspectos cronoculturales, etc. Esto da lugar a que las características de los productos de talla varíen a lo largo del proceso de reducción. Bajo diferentes términos es común encontrar alusiones a un método de talla intermedio entre el discoide y el Levallois, presentando una amplia distribución geográfica y cronológica a lo largo del Paleolítico Medio y el Paleolítico Medio inicial. La concepción de este método de talla, referido en el presente documento como Discoide Jerárquico, posee características intermedias entre el Levallois (jerarquización de superficies no intercambiables o un plano de talla paralelo a la intersección de ambas superficies) y el discoide (ausencia de preparación de talones, planos de talla secantes en la fase inicial de talla), surgiendo la duda de hasta qué se pueden diferenciar los productos de lascado de los tres métodos y sobre la direccionalidad de las confusiones.
El presente trabajo emplea un conjunto experimental de lascas procedentes de los tres métodos de talla (77 del método de talla discoide, 73 del Levallois y 72 del Discoide Jerárquico). Sobre este conjunto experimental de lascas se realiza un análisis métrico y de atributos, y sobre los datos procedentes de este análisis se entrenan diez algoritmos de aprendizaje automático con el objetivo de determinar hasta qué punto es posible diferenciar el método de talla. Para evaluar los algoritmos de aprendizaje automático se tiene en cuenta la precisión general de los modelos, pero también los efectos del uso de umbrales de probabilidad en la identificación de los métodos de talla. El uso de umbrales de probabilidad permite optimizar el ratio de positivos verdaderos y positivos falsos para cada umbral de decisión y de ahí extraer el “área bajo la curva” (AUC en inglés) como valor de avaluación de un modelo.
De los diez algoritmos de aprendizaje automático, una máquina de vector soporte con kernel polinomial presenta los mejores resultados en la identificación de los tres métodos de talla, proporcionando unos resultados excelentes a la hora de diferenciar entre los tres métodos a nivel general (0.667 precisión, 0.824 AUC). Considerando individualmente cada método de talla, los resultados subrayan el carácter singular de las lascas procedentes de secuencias de reducción Levallois ya que obtienen una identificación excepcionalmente buena (AUC de 0.91), siendo su procedencia raramente atribuida a cualquiera de los otros dos métodos. La confusión entre productos procedentes de secuencias de talla discoide y el Discoide Jerárquico es más común, aunque se alcanza una identificación excelente en el caso de los productos procedentes de reducciones discoides (AUC de 0.82) y una identificación aceptable en el caso los productos procedentes del Discoide Jerárquico (AUC de 0.73).
Estos resultados muestran el potencial de combinar modelos de aprendizaje automático con análisis de atributos sobre lascas para la identificación de métodos de talla. Su uso puede servir de gran ayuda en la identificación de métodos de talla en lascas. Sin embargo, su uso requiere de una evaluación previa de los conjuntos líticos para determinar posibles métodos de talla existentes, uso diferencial de las materias primas, y evaluación de las cadenas operativas presentes.
Palabras clave: tecnología lítica; arqueología experimental; Levallois; Discoid; Paleolítico Medio; Aprendizaje Automático
The Middle Palaeolithic in western Europe is characterised by the increase in and diversification of prepared core knapping methods, resulting in flake-dominated assemblages (Kuhn 2013). These flake-dominated assemblages are the result of a wide number of production methods including Levallois (Boëda 1994; Boëda 1995b; Boëda et al. 1990), discoid (Boëda 1993; Boëda 1995a), the système par surface de débitage alterné or SSDA (Forestier 1993; Ohel et al. 1979), Quina (Bourguignon 1996; Bourguignon 1997), different laminar production systems (Boëda 1990; Révillon & Tuffreau 1994), and the Kombewa (Newcomer & Hivernel-Guerre 1974; Tixier & Turq 1999) among several others. This abundance of different production methods results in a highly diversified material culture in which flakes exhibit great morphological variability. Flakes often retain morphologies and attributes characteristic of the knapping method used to detach them, facilitating the identification of those methods. However, flakes also often present overlapping attributes and morphologies as a result of the high internal variability of the methods used and the fact that flakes with similar functional properties can be produced via different methods Kuhn (2013).Due to their extensive geographical and chronological distribution, the Levallois and discoid constitute important sources of cultural variability in the Middle Palaeolithic of western Europe.
Boëda (1994; 1995b) establishes six characteristics that define the Levallois knapping strategy from a technological perspective:
Depending on the organisation of the debitage surface, Levallois cores are usually classified as a preferential method (in which a single predetermined Levallois flake is obtained from the debitage surface) or as recurrent methods (in which several predetermined flakes are produced from the debitage surface) with removals being either unidirectional, bidirectional or centripetal (Boëda 1995b; Boëda et al. 1990; Delagnes 1995; Delagnes & Meignen 2006).
Because of its early recognition in the XIX century (Mortillet 1885: pp.:235–236), its association with the cognitive abilities of planning and predetermination (Boëda 1994; Pelegrin 2009), and its use for the definition of cultural facies (Bordes 1961a; Bordes 1953) and lithic technocomplexes (Delagnes et al. 2007; Faivre et al. 2017), Levallois flaking technology is considered a trademark of the Middle Palaeolithic. The emergence of the Levallois method has been observed from MIS 12 to MIS 9, with several sites yielding artefacts characteristic of this knapping strategy (Carmignani et al. 2017; Hérisson et al. 2016; Moncel et al. 2020; Soriano & Villa 2017; White & Ashton 2003). However, the Levallois is clearly generalised and identified from MIS 8 onwards, covering an extensive geographical area throughout western Europe (Delagnes et al. 2007; Delagnes & Meignen 2006; Faivre et al. 2017; Geneste 1990).The extensive geographical area and long temporal span of the Levallois creates additional layers of variability which can result from raw material constraints, synchronic variability as a result of different site functionalities, chronological trends in the development of methods, or shifts in the technological organisation of groups. It is also important to highlight the explicit recognition of Levallois cores after MIS 8, while a multitude of terms were employed to define previous hierarchical knapping strategies and their possible coexistence with Acheulean technocomplexes (Moncel et al. 2020; Santonja et al. 2016; Hérisson et al. 2016; Rosenberg-Yefet et al. 2022; White & Ashton 2003; Scott & Ashton 2011).
Boëda (1993; 1994; 1995a), also establishes six technological criteria defining the discoid method:
Technological analyses of Middle Palaeolithic assemblages have gradually led to the identification of a variability of modalities within discoid core knapping (Bourguignon & Turq 2003; Locht 2003; Terradas 2003; Locht 2003). This has resulted in sensu stricto and a sensu lato conceptualisations of the discoid knapping system (Faivre et al. 2017; Mourre 2003; Thiébaut 2013). The sensu stricto hmethod closely corresponds to the definition established by Boëda (1993) described above, for which core edge flakes and pseudo-Levallois points are the most common products. The sensu lato discoid encompasses a greater range of products (although centripetal flakes are more common) as a result of higher variability in the organisation of percussion and exploitation surfaces (Terradas 2003).
One of the variants of the sensu lato discoid conceptualisation resembles the Levallois knapping strategy (Figure 1). Several common characteristics have been defined for this method:
Schematic representation of the knapping methods, surfaces and platform preparation
Strategies evidenced at several sites fit into the discoid knapping variation described above, and its resemblance to the Levallois has been previously noted in several Middle and early Middle Palaeolithic assemblages (Casanova i Martí et al. 2009; 2014; Jaubert 1993; Peresani 1998; Slimak 1998; 2003; Soriano & Villa 2017). However, it is important to consider that a variety of different terms have been employed due to the broad geographical and chronological span of the method. For Middle Palaeolithic sites, the identification of this method usually focuses on its shared features with the Levallois and discoid methods and thus, its intermediate nature.
Jaubert (1993), at Mauran, noted the hierarchical nature of the production system and its resemblance to exhausted recurrent centripetal Levallois cores. However, he pointed out that the secant planes of detachment were not as parallel as in the Levallois. Slimak (1998; 2003), at Champ Grand, also noted the similarities of residual cores with recurrent centripetal Levallois debitage. At Estret de Tragó, Casanova i Martí et al. (2009; 2014)) noted the presence of products and knapping methods which shared features of the Levallois and discoid methods, and proposed including the hierarchical discoid and Levallois recurrent centripetal strategies within a single hierarchical bifacial centripetal class. Peresani (1998) at Fumane cave, documented the presence of debitage products with features (reduced thickness, debitage angle, and centripetal organisation of scars also subparallel to the ventral surface) which would correspond to parallel-plane debitage. Baena et al., (2005) indicated the presence of the hierarchical discoid method throughout the Esquilleu cave sequence as a secondary and primary knapping method.